Detecting age-related macular degeneration (AMD) biomarker images using MFCC and texture features

Age-related macular degeneration (AMD) is the leading cause of irreversible vision loss in older individuals. Clinically, ophthalmologists visually inspect optical coherence tomography (OCT) volumes to diagnose the stage of AMD based on well-known biomarkers. An early characteristic of AMD is drusen, which appears as yellowish deposits under the retina. AMD is mainly categorized into two types: dry AMD (non-neovascular) and wet AMD (neovascular). Given the large number of OCT images in an individual volume, an efficacious computer-aided detection system can reduce the workload for ophthalmologists by automatically detecting biomarkers in the relevant images. Because the shape of the RPE is critical in defining the pathological changes caused by wet and dry AMD, we propose a novel approach to describe the RPE shape using Mel Frequency Cepstral Coefficients (MFCC). Our previous work indicates that Haralick texture features have the ability to distinguish drusen from healthy tissue on color photography, therefore, we also investigated Haralick texture features extracted from the region between Inner Limiting Membrane (ILM) and Bruchs Membrane (BM) layers in this study. We achieved a mean accuracy, sensitivity with respect to AMD image and specificity with respect to healthy image of 89.68%, 89.26% and 90.12% on testing sets and 69.22%, 67.40%, and 75.56% on new patient validation sets, respectively. Our binary classification results indicate that MFCC are uniquely suited for producing generalizable results to automatically detect AMD biomarker images.